lovart3/app/api/canvas/text-detect/route.ts

import { NextResponse } from "next/server";
import { z } from "zod";
import { generateText } from "ai";
import { auth } from "@/app/(auth)/auth";
import { getEnabledAiModelsWithProvider } from "@/lib/db/queries";
import { getLanguageModelById } from "@/lib/ai/dynamic-provider";

export const runtime = "nodejs";
export const maxDuration = 60;

const requestSchema = z.object({
  imageUrl: z.string().min(1, "图片地址不能为空"),
});

interface OcrItem {
  id: string;
  text: string;
  confidence?: number;
}

function pickVisionModel(models: Awaited<ReturnType<typeof getEnabledAiModelsWithProvider>>) {
  const normalized = models
    .filter((item) => item.model.supportsVision)
    .sort((a, b) => {
      const aScore =
        (a.model.modelId?.toLowerCase().includes("qwen3-vl") ? 100 : 0) +
        (a.model.displayName?.toLowerCase().includes("qwen3-vl") ? 100 : 0) +
        (a.provider.type === "alibaba" ? 20 : 0) +
        (a.model.isDefault ? 5 : 0);
      const bScore =
        (b.model.modelId?.toLowerCase().includes("qwen3-vl") ? 100 : 0) +
        (b.model.displayName?.toLowerCase().includes("qwen3-vl") ? 100 : 0) +
        (b.provider.type === "alibaba" ? 20 : 0) +
        (b.model.isDefault ? 5 : 0);
      return bScore - aScore;
    });

  return normalized[0];
}

function parseJsonText(raw: string): any | null {
  if (!raw) return null;

  const trimmed = raw.trim();
  try {
    return JSON.parse(trimmed);
  } catch {
    // continue
  }

  const markdownJsonMatch = trimmed.match(/```json\s*([\s\S]*?)\s*```/i);
  if (markdownJsonMatch?.[1]) {
    try {
      return JSON.parse(markdownJsonMatch[1]);
    } catch {
      // continue
    }
  }

  const objectMatch = trimmed.match(/\{[\s\S]*\}/);
  if (objectMatch?.[0]) {
    try {
      return JSON.parse(objectMatch[0]);
    } catch {
      // continue
    }
  }

  return null;
}

function sanitizeItems(items: unknown[]): OcrItem[] {
  const result: OcrItem[] = [];

  for (let index = 0; index < items.length; index += 1) {
    const item = items[index] as Record<string, unknown>;
    const rawText = typeof item?.text === "string" ? item.text : "";
    const text = rawText.replace(/\s+/g, " ").trim();
    if (!text) continue;

    const confidenceValue =
      typeof item?.confidence === "number" && Number.isFinite(item.confidence)
        ? Math.min(Math.max(item.confidence, 0), 1)
        : undefined;

    result.push({
      id: String(item?.id || index + 1),
      text,
      confidence: confidenceValue,
    });
  }

  return result.slice(0, 50);
}

function fallbackItemsFromRaw(raw: string): OcrItem[] {
  const lines = raw
    .split("\n")
    .map((line) => line.replace(/^[-*\d.\s]+/, "").trim())
    .filter(Boolean)
    .slice(0, 20);

  return lines.map((text, index) => ({
    id: String(index + 1),
    text,
  }));
}

export async function POST(request: Request) {
  try {
    const session = await auth();
    if (!session?.user?.id) {
      return NextResponse.json({ success: false, error: "未登录" }, { status: 401 });
    }

    const body = await request.json();
    const parsed = requestSchema.safeParse(body);
    if (!parsed.success) {
      return NextResponse.json({ success: false, error: parsed.error.errors[0]?.message || "参数错误" }, { status: 400 });
    }

    const imageUrl = parsed.data.imageUrl.trim();
    if (!/^https?:\/\//i.test(imageUrl)) {
      return NextResponse.json({ success: false, error: "仅支持 URL 图片，不允许 base64 传输" }, { status: 400 });
    }

    const models = await getEnabledAiModelsWithProvider();
    const selected = pickVisionModel(models);

    if (!selected) {
      return NextResponse.json({ success: false, error: "未配置可用的视觉模型" }, { status: 500 });
    }

    const model = await getLanguageModelById(selected.model.id);
    if (!model) {
      return NextResponse.json({ success: false, error: "视觉模型加载失败" }, { status: 500 });
    }

    const prompt = [
      "你是图片文字识别助手。",
      "请识别图片中可见文字并按阅读顺序输出 JSON。",
      "只返回 JSON，不要额外说明。",
      "格式：{\"items\":[{\"id\":\"1\",\"text\":\"示例\",\"confidence\":0.95}]}",
      "如果没有识别到文字，返回 {\"items\":[]}。",
    ].join("\n");

    const { text } = await generateText({
      model,
      messages: [
        {
          role: "user",
          content: [
            { type: "text", text: prompt },
            { type: "image", image: imageUrl },
          ],
        },
      ],
      temperature: 0.1,
    });

    const parsedJson = parseJsonText(text || "");
    let items: OcrItem[] = [];

    if (parsedJson && Array.isArray(parsedJson.items)) {
      items = sanitizeItems(parsedJson.items);
    }

    if (items.length === 0 && text) {
      items = fallbackItemsFromRaw(text);
    }

    return NextResponse.json({
      success: true,
      data: {
        items,
        model: selected.model.displayName,
      },
    });
  } catch (error) {
    console.error("[Canvas Text Detect] Error:", error);
    return NextResponse.json(
      {
        success: false,
        error: error instanceof Error ? error.message : "文字识别失败",
      },
      { status: 500 },
    );
  }
}